Protein Structure Prediction with NVIDIA and Claude Science
A new era in AI-powered protein structure analysis begins.
1Revolution in Protein Structure Prediction
The NVIDIA BioNeMo Agent Toolkit offers a groundbreaking set of tools for protein structure prediction. It consolidates over a decade of BioNeMo life sciences models, libraries, and workflows into a single platform. The toolkit enables protein structure prediction using multiple-sequence alignment (MSA) and multiple folding models. OpenFold3 and Boltz-2 models stand out with high interface confidence scores (iPTM 0.85 and 0.82 respectively) in benchmark tests.
2Claude Science Integration
Through collaboration between NVIDIA and Anthropic, the BioNeMo Agent Toolkit has been integrated into the Claude Science platform. This integration allows AI agents to discover, launch, and directly call BioNeMo NIM microservices. The system can manage complex scientific workflows in specialized domains. Internal benchmarks show it improves task accuracy from 60% to 100% and roughly doubles token efficiency.
3Technical Requirements
Running this workflow requires a workstation or cloud machine with NVIDIA L40S or H100 GPU. The system needs approximately 700GB of storage: ~490GB for the UniRef30 database, plus 30-40GB total for Boltz-2 and OpenFold3 containers. While Claude Science runs in a sandbox by default, running BioNeMo NIM microservices requires endpoints that can access local or remote GPU resources.
4Setup and Configuration
Users must first import BioNeMo Agent Toolkit skills from GitHub, add an NVIDIA API key, and connect to local endpoints. Then they need to download and launch local BioNeMo NIM containers for msa-search, openfold3 and boltz2 microservices. Each microservice requires health checks and endpoint approval. This process involves downloading containers for each microservice.
5Case Study: Seh1 Protein
The tutorial examines the Seh1 protein and a proposed partner (C1HCX1) from the Paracoccidioides lutzii fungus. Seh1 and Mio-family proteins participate in conserved nutrient-sensing machinery. Structures from other organisms show that Mio can complete Seh1's otherwise open β-propeller, making this pair particularly useful for testing MSA-supported complex prediction.
6Performance Comparison
In benchmarks comparing Seh1 monomer and Seh1-C1HCX1 heteromer predictions, both OpenFold3 and Boltz-2 achieved high interface confidence (iPTM 0.85 and 0.82) with MSA input. Without MSA, scores dropped dramatically to 0.14 and 0.19, demonstrating that evolutionary alignment is critical for interface prediction. Structural superposition showed that with a partner present, Seh1's fold is completed rather than remodeled.
7AI in Scientific Research
Agentic AI is transforming research methodologies. AI scientists can now read papers, propose hypotheses, call models, and determine which experiments to prioritize next. Having proven their value in software engineering where coding agents write, test, and ship production code, these systems are now tackling more demanding and iterative scientific research processes where evidence evaluation and hypothesis refinement are continuous.
8Domain-Specific Tools
Scientific problems often require specialized tools - for protein folding or molecular characterization, for example. Orchestrating these tools presents challenges as similar packages may have vastly different environment requirements or APIs. While a general-purpose agent might recognize that a task requires protein folding or molecular docking, it wouldn't inherently know which model to run, how to format the request, or which parameters matter most.
9Advantages of BioNeMo Agent Toolkit
The NVIDIA BioNeMo Agent Toolkit bridges this gap by packaging over a decade of NVIDIA BioNeMo life sciences models, libraries, and workflows into agent-callable skills for biology, chemistry, genomics, and drug discovery. Designed to work with any agent framework, it enables management of complex scientific workflows requiring specialized domain expertise.
10Next Steps
Researchers can explore the BioNeMo Agent Toolkit on GitHub to apply this workflow to other monomers and complexes. Powered by NVIDIA Nemotron, this toolkit is shaping the future of AI-assisted scientific research. However, users should remember that AI-generated content may summarize information incompletely, so important findings should always be verified independently.
11New Era in Scientific Discovery
The integration of NVIDIA BioNeMo with Claude Science heralds a new era in protein structure prediction and drug discovery. This technology enables researchers to obtain faster, more accurate results while increasing the efficiency of scientific processes. It appears particularly promising for understanding rare diseases and complex biological processes, potentially accelerating breakthroughs in these challenging areas.
Original article: Nvidia Developer Blog — NVIDIA ve Claude Science ile Protein Yapısı Tahmini